2025
Graph-Based Uncertainty Modeling and Multimodal Fusion for Salient Object Detection
Yuqi Xiong, Wuzhen Shi, Yang Wen, Ruhan Liu
International Conference on Neural Information Processing (ICONIP) 2025
This project focuses on salient object detection in complex scenes, where we developed a dynamic uncertainty-aware graph convolution module to enhance the detection of fine structures and edge regions, addressing detail loss and blurred boundaries. We further designed a multimodal fusion strategy to adaptively integrate RGB, depth, and edge features, enabling cross-modal semantic complementarity and consistency. Experiments on standard datasets demonstrated improved edge clarity and robustness against complex backgrounds.
Graph-Based Uncertainty Modeling and Multimodal Fusion for Salient Object Detection
Yuqi Xiong, Wuzhen Shi, Yang Wen, Ruhan Liu
International Conference on Neural Information Processing (ICONIP) 2025
This project focuses on salient object detection in complex scenes, where we developed a dynamic uncertainty-aware graph convolution module to enhance the detection of fine structures and edge regions, addressing detail loss and blurred boundaries. We further designed a multimodal fusion strategy to adaptively integrate RGB, depth, and edge features, enabling cross-modal semantic complementarity and consistency. Experiments on standard datasets demonstrated improved edge clarity and robustness against complex backgrounds.
Non-Stationary Time Series Forecasting Based on Fourier Analysis and Cross Attention Mechanism
Yuqi Xiong, Yang Wen
International Joint Conference on Neural Networks (IJCNN) 2025
Proposed AEFIN, a novel framework for non-stationary time series forecasting that combines Fourier Transform-based decomposition to separate stationary and non-stationary components, cross-attention mechanisms for enhanced information sharing, and Fourier analysis for capturing periodic patterns. Designed a custom loss function that integrates MSE and MAE across time- and frequency-domain components to improve model robustness. The proposed framework achieved SOTA performance on most benchmark datasets, consistently improving MSE and MAE when integrated with mainstream Transformer- and CNN-based forecasting models.
Non-Stationary Time Series Forecasting Based on Fourier Analysis and Cross Attention Mechanism
Yuqi Xiong, Yang Wen
International Joint Conference on Neural Networks (IJCNN) 2025
Proposed AEFIN, a novel framework for non-stationary time series forecasting that combines Fourier Transform-based decomposition to separate stationary and non-stationary components, cross-attention mechanisms for enhanced information sharing, and Fourier analysis for capturing periodic patterns. Designed a custom loss function that integrates MSE and MAE across time- and frequency-domain components to improve model robustness. The proposed framework achieved SOTA performance on most benchmark datasets, consistently improving MSE and MAE when integrated with mainstream Transformer- and CNN-based forecasting models.